Solving Multi-Parameter Optical Spectroscopy Inverse Problems Using Transfer Learning

Abstract

Spectroscopic analysis of multicomponent media is commonly formulated as an inverse problem (IP), where the concentrations of individual components are reconstructed from measured spectra, whose shape is determined by the composition of the medium.
Data-driven machine learning approaches provide an effective way to solve such IPs. However, their practical application is constrained by the need for large experimental datasets.
This study considers the IP of determining the concentrations of metal ions in aqueous solutions from excitation-emission matrices of carbon dots (synthesized by the hydrothermal method from citric acid and ethylenediamine) introduced into the solution. The source problem involves determination of six ions (Cu²⁺, Ni²⁺, Al³⁺, Co²⁺, Cr³⁺ cations and NO₃⁻ anion), while
the target problem extends the analyzed system by an additional Pb²⁺ ion.
To reduce experimental and computational costs associated with this transition,
transfer learning (TL) techniques are investigated for knowledge transfer from the six-parameter to the seven-parameter (7P) IP. Particular attention is paid to source-target domain data mixing strategies, including different dynamic weighting schemes for source and target domain data during model optimization. Their performance is compared on independent test dataset of the target 7P problem with conventional fine-tuning and training models from scratch.
The results demonstrate that appropriate TL strategies enable accurate solution of the expanded spectroscopy IP while reducing the amount of target-domain experimental data and computational resources required for model development.
This study has been conducted at the expense of Russian Science Foundation grant no. 24-11-00266, https://rscf.ru/en/project/24-11-00266/.

Speaker

Artem Guskov
Faculty of Physics, Moscow State University, Moscow, Russia
Russia

Discussion

Ask question